The AI Generated Art Thread

Whether this can be considered derivative art or not (and how much it depends on each particular image) has not been tested legally afaik and it’s definitely a gray zone with some outputs. What we think about it’s derivative nature doesn’t matter, other than being a possible line of defense towards a claim (a line of defense that could be accepted or not). The law as written now makes a claim at the very least viable in some circumstances.

So you think there’s a legal difference depending on what the user prompt was? What exactly makes it legally different than if the user prompt didn’t include the reference to the copyrighted image?

Think of it as hiring an art outsourcer. I say “I want a superhero sticking to a wall in a heroic pose” without ever referencing Spiderman. The artist gives me an Spiderman image. I can’t legally use that image anyway, even if my “prompt” to my source was innocuous, because my source has been unduly influenced by copyrighted work and her work could be considered derivative.

It is a huge can of worms until we have laws covering use cases.

Love the artwork here - I’ve used Midjourney a bit as well, and this is my best result so far.

Those are very, very nice!

Midjourney does seem to create the most, for want of a better word, natural looking images of all the “Dall-E Mini” tools, though none of them come close (at least without a lot of work) to the full fat AIs.

Midjourney dataset seems more art/paint oriented, while Dalle has more normal photos stuff.

Well, to be fair, your products are exquisitely good!

Art is subjective. I look at the stuff in my old D&D or Traveller books from yore and compare to the output I got and yeah, I do think that cartoon image of a robot with some kind of instrument is better than the scribbles I find in those published works.

If you’re looking for very specific output like the Star Wars characters you posted, then yes it’s going to take a few tries. You said 40-ish for those? How long did that take? How does that compare to commissioning art from someone? Was that easier or harder?

My point is, I don’t find Midjourney hard to use at all; certainly easier than going back and forth with a real person trying to produce something. And it’s only going to get easier, faster, and better.

I guess I should try for myself, but how easy is it to use for a very non-artistic person?
I think it’s still at the stage of a tool for artists, because there’s still all the issues of coherence with a larger work, or corporate image, or… but I am not one who notices or cares about such things, and, well, at some level of quality, most people eventually won’t either.

You join the Midjourney Discord, then type “/imagine” then whatever you want Midjourney to paint.

Again, if you want a very specific thing it may take a few tries, but it’s not hard at all.

Well, yes, I googled. I meant how easy is it to create something that would be used commercially?

Those last three are hauntingly beautiful

Wait, what? Is this accurate?
Derivative works can be copyrighted on their own, and I’m pretty sure you can in fact use them commercially, but maybe I’m mistaken.

I mean, of course you can, if you own copyright (or a license) to the original copyrighted work.

I should have said you can’t commercialize a derivative work of copyrighted material (for example, sell a fan written Harry Potter novel) without a license from the copyright holder.

I think the reason you can’t write a fan written Harry Potter novel is because the character itself is copyrighted. Thus, a new story with Harry Potter would not be a derivative work, but would instead be directly violating the copyright on the character of Harry Potter.

If you make a picture, and then I make a derivative work of that picture that meets the requirements of it being a derivative work (meaning that it’s sufficiently different), I believe I actually can copyright it myself. I’m pretty sure that this is laid out explicitly in US copyright law.

You can copyright the changes you did to the original work, but since it’s a derivative work you need the right to commercialize if that stems from the original copyright.

There have been cases of plagiarism where a work that follows too closely a plot of another work (even changing locations and character names) has been deemed in breach of copyright by tribunals.

Hmm, I didn’t realize that the original copyright holder also automatically had rights over all the derivative works.

It looks like I was confusing it with “transformative” works.
https://foundrylawgroup.com/copyright-copywrong-what-are-derivative-and-transformative-works/

In some of these AI art cases, I think you’d have a strong case that the resulting work would be transformative, rather than derivative.

Of course, that’s why I say it’s a very gray area legally. It’s going to take some borderline case to establish some legal guidelines. I personally think the vast (very vast) majority of AI generated works are going to be fine. But the line of what’s transformative is quite muddy already.

My main concern is how much user prompt (intention) matters in both adjudicating whether there’s a copyright breach, and in the case there is, who breached.

Of course if you do a prompt of “Spiderman jumping” and you get sued because it’s clearly Spiderman, it’s clearly your fault. Don’t prompt anything that is copyrighted and expect to be magically ok because “it’s AI”.

But the most fascinating case is that of a work could possible be derivative (in the sense of being clearly recognizable as a copyright violation) while the input (user prompt) does not reference the copyrighted material. Spiderman is widely recognizable, but what if some of those portraits linked above are too close to a copyrighted character (minor) the AI has been trained upon?

Legally it’s unclear. The algorithm is using the original copyrighted work to give an output and the output is clearly recognizable, yet the user never knew. Is it a copyright violation? Very likely (I think it could be one even if the original work was not part of the training set). But who’s liable? The company that made the AI and incorrectly trained on copyrighted material or the user that used it unknowingly?

I find this kind of stuff fascinating.

Yeah, on some level, that idea that it’s “trained on copyrighted material” doesn’t matter much to me, because you can say that every human artist is “trained on copyrighted material”.

The difference is that it’s trackable. That is, in a trial you can show the exact image the AI was trained upon, so a lot of frequent outs in such cases (I never even read the book is a common one) won’t fly.

It’s also a mechanical process without authorship nor copyright protection (so far, I think that will change, since it’s quite ridiculous) and that will give more weight to the original copyright in case of a dispute.

In my opinion these systems should be trained on non copyrighted images until the law becomes clearer (which would make them much less useful, so it’s not practical), or the output never used commercially (which is the safer thing to do for now). For most cases it’s going to be fine, but someone might get unlucky at some point. It’s definitely a legal liability.

Do we know that these are trained on copyrighted material? I would be surprised if the lawyers in these organizations weren’t very careful about the licenses on the images that went into the training data.

I suppose it’s more complex than that. An image might be licensed to be used in cases X, Y, and Z which cover training an ML model, but not for (say) commercial re-sale, but what does that mean for the output of the ML model when it comes to commercial re-sale?

Clearly, since they understand copyrighted stuff. It couldn’t understand “Star Wars”, “The night King”, “Godzilla” etc… if not.

It could be trained on fan stuff that breaks copyright. That is, fan images available online on the internet that are breaching copyright (but nobody cares). And even those have copyright. Basically, every artistic creation has copyright unless otherwise stated by the creator.

No way these use only public domain training sets (it’s impossible because of the aforementioned understanding of copyrighted terms).